Remote Sensing

Abstract: Hyperspectral images (HSI) are corrupted by a combination of Gaussian and impulse noise. Successful denoising of HSI data increases the accuracy of high-level vision operations like classification, target tracking and land-cover problem. On the one hand, the traditional approach of handling the denoising problem using maximum a posteriori (MAP) criterion is often restricted by the time-consuming iterative optimization process and design of hand-crafted priors to obtain an optimal result. On the other hand, the discriminative learning-based approaches offer fast inference speed over a trained model; but are highly sensitive to the noise level used for training. A discriminative model trained with a loss function which does not accord with the Bayesian degradation process often leads to sub-optimal results. In this paper, we design the training paradigm emphasizing the role of loss functions in neural network; similar to as observed in model-based optimization methods. Further, Bayesian motivated loss functions also act as priors to constrain the solution space to the types of noise observed in hyperspectral image acquisition process. As a result, loss functions derived in Bayesian setting and employed in neural network training boosts the denoising performance. Extensive analysis and experimental results on synthetically corrupted and real hyperspectral datasets suggest the potential applicability of the proposed technique under a wide range of homogeneous and heterogeneous noisy settings. Classification results over the real dataset and associated metrics like kappa coefficient and overall accuracy used as the task-based evaluators further support our hypothesis made in the proposed methodology. (Online)

Abstract: In this letter, we present a deep-learning-based methodology for recovering hyperspectral images (HSIs) distorted by Gaussian and impulsive noise. This work makes the following contribution: To begin with, the Wasserstein Generative Adversarial Network (WGAN) is used to mitigate the effects of vanishing gradient and mode collapse that can occur when training a vanilla GAN. Secondly, data are passed via three distinct pathways in a parallel ensemble to promote multiscale feature extraction. Normal and multiscale dilated 3D convolutions are utilized to train the model in each pair of parallel paths. Thirdly, features are recovered following data permutation across three different spatial planes (viz. xy; yz, and xz planes) and after passing through parallel convolutional blocks; to promote spatio-spectral similarity within and across the different layers of the HSI data. Fourthly, by adopting Structural Similarity (SSIM) as the content loss, the issue of loss in resolution encountered during adversarial training is mitigated. Finally, the incorporation of 3D depth-wise separable convolution and batch re-normalization (BRN) solves the major issue of computational burden encountered while processing HSI data. Extensive experimental evaluation on synthetically corrupted data and real HSI data (obtained from real hyperspectral sensors) under various degradation conditions suggests that the aforementioned denoising approach could be used in real time. (Online) (Supplementary Text)